A roof greening trend identification and prediction method based on deep time series modeling

Abstract

This invention discloses a method for identifying and predicting rooftop greening trends based on deep time-series modeling, relating to the fields of remote sensing technology and greening analysis. By integrating a three-stage analysis system of self-supervised encoder, mutation detection, and Transformer predictor, the method utilizes the SatMAE-NDVI model to achieve trend discrimination and mutation point identification of building NDVI time series. Based on this, the method employs a Transformer structure to predict and classify the greening evolution trend for the next three years. This method can achieve automatic analysis and forward-looking prediction of the entire process of rooftop greening evolution at the building scale, and is suitable for the refined management and dynamic optimization of green infrastructure in complex urban environments.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing technology and greening analysis, and more specifically, to a method for identifying and predicting rooftop greening trends based on deep time-series modeling. Background Technology

[0002] In the construction of urban green infrastructure, rooftop greening has become an important way to promote the growth of green space in high-density urban areas due to its multiple ecological functions, such as increasing urban green space, mitigating the heat island effect, and improving the microclimate. However, existing rooftop greening monitoring methods still have two technical shortcomings: First, most methods rely on single-temporal remote sensing images and lack systematic utilization of multi-temporal NDVI data, making it difficult to accurately reflect the temporal changes of rooftop greening, especially at the building scale, where the continuity of temporal information is poor and the ability to identify trends and abrupt changes is limited. Second, under the conditions of dense urban buildings and complex roof shapes, the spatial matching accuracy between building outlines and remote sensing pixels is low, and the registration error is significant. In addition, problems such as cloud shadow interference and image distortion make the temporal expression of greening status unstable, making it difficult to support refined dynamic analysis and classification judgment.

[0003] Furthermore, in high-density cities, rooftop greening exhibits diverse evolutionary paths and varying rates of change, making it difficult for traditional statistical methods to accurately identify its temporal trends and predict future directions. Specific challenges include the coexistence of long-term trends and local abrupt changes in NDVI time-series data, and data noise and missing information caused by factors such as building renovation and shading, affecting the stability and consistency of trend identification. Simultaneously, existing methods struggle to organically combine historical trend identification with future change prediction, lacking integrated dynamic modeling capabilities. Therefore, there is an urgent need to propose a trend-abrupt change-prediction joint modeling method that integrates deep learning models to achieve systematic identification of the state of rooftop greening at the building scale and classification of future trends. Summary of the Invention

[0004] To address the challenge of dynamically identifying and predicting rooftop greening trends in high-density cities, this invention overcomes the shortcomings of existing technologies by proposing a method for identifying and predicting rooftop greening trends based on deep time-series modeling. This method constructs a three-stage analysis system integrating a self-supervised encoder, mutation detection, and a Transformer predictor. It utilizes the SatMAE-NDVI model to discriminate trends and identify mutation points in the NDVI time series. Furthermore, it employs a Transformer structure to predict and classify the greening evolution trend over the next three years. This approach enables automatic analysis and forward-looking prediction of the entire rooftop greening evolution process at the building scale, making it suitable for the refined management and dynamic optimization of green infrastructure in complex urban environments.

[0005] The first aspect of this invention provides a method for identifying and predicting rooftop greening trends based on deep time-series modeling, comprising:

[0006] S1: Obtain Sentinel-2 NDVI time-series images for a preset number of years, extract building outlines and construction years using OpenStreetMap, and construct the NDVI time series for each building;

[0007] S2: The SatMAE-NDVI autoencoder model is used for self-supervised pre-training. Temporal features are extracted through the mask reconstruction mechanism, and a multi-class classifier is used to determine the overall trend direction of the model output. In each NDVI time series, the reconstruction error sequence and attention weight are output, and the potential mutation point is detected. The NDVI time series is divided into two segments according to the mutation point, and the segments are input into the lightweight regressor to predict the slope value of each segment.

[0008] S3: Based on the relationship between the slope values ​​before and after the mutation point and the overall trend direction, the greening evolution of each NDVI time series is classified into multiple modes.

[0009] S4: After completing the classification of greening evolution, a predictor with Transformer as the core is constructed. The NDVI time series after classification is input for prediction, and the NDVI prediction value for the next three years is output. The predictor uses multi-layer temporal location coding and multi-head self-attention to capture long-term dependencies for sequence prediction.

[0010] In this solution, S1 specifically refers to:

[0011] Sentinel-2 L2A multi-temporal remote sensing images from 2018 to 2024 were retrieved using GEE. All remote sensing images were preprocessed with cloud masking and radiometric normalization. The 95th percentile NDVI value of the year was extracted for each pixel to obtain the pixel-level NDVI.

[0012] Using OpenStreetMap building vector outlines and ArcGIS Zonal Statistics, pixel NDVI is aggregated by building unit to form an annual sequence of length 7 in the time dimension. Combined with the year the building was first constructed Eliminate Observed values;

[0013] Based on the building outline and the year of construction, an annual NDVI time series is generated for each building.

[0014] In this solution, S2 specifically refers to:

[0015] The NDVI time series data was imported into the SatMAE-NDVI autoencoder model for self-supervised pre-training.

[0016] By employing a mask reconstruction method to learn the latent temporal structure of NDVI sequences, a three-classifier is used to determine the overall trend direction of greening. In each NDVI time series, the mean is calculated based on the sliding window of the model, and the reconstruction error and the location of the mutation point are calculated. A lightweight regressor is used to output the slope values ​​of the segments before and after the mutation point.

[0017] In this solution, the overall trend direction determination in S2 is specifically as follows:

[0018] NDVI time series The input is fed into a SatMAE-NDVI autoencoder model, which is pre-trained on multiple NDVI time series using a mask reconstruction task to obtain latent vectors that can simultaneously represent multi-year trends and seasonal rhythms. During the fine-tuning phase, a three-class classifier is constructed. right Perform discrimination and output strictly defined trend labels. The three-category representation is as follows:

[0019] ,

[0020] in, Indicates the overall trend classification. This indicates three classification results.

[0021] In this solution, S2 further includes:

[0022] For each building sequence, re-perform mask reconstruction of the SatMAE-NDVI autoencoder model and calculate the time-by-time reconstruction error. Calculate the mean within a sliding window of length w. ,when When the value exceeds the mean of the entire sequence plus three standard deviations, that moment is fixed as the unique mutation point. ,by The building sequence is divided into two parts, with the first part serving as the boundary. With the latter part And input them into the lightweight slope regressor respectively. Output and The details are as follows:

[0023] ,

[0024] ,

[0025] in, and All of these are output slope values.

[0026] In this solution, S3 specifically refers to:

[0027] Based on the sign relationship between the overall trend label and the slopes of the two segments, the greening evolution of the building sequence is classified into seven typical trend types, as follows:

[0028] Continued to rise: , , ;

[0029] Continued decline: , , ;

[0030] It first rises and then falls, but the overall trend is upward: , , ;

[0031] It first rose and then fell, resulting in an overall decline. , , ;

[0032] First decrease, then increase, overall rise: , , ;

[0033] First decrease, then increase, overall decrease: , , ;

[0034] No clear trend: The overall trend is not significant.

[0035] In this solution, S4 specifically refers to:

[0036] Build a Transformer-based predictor;

[0037] After obtaining the classification results of the seven typical S3 patterns, the complete NDVI time series is input into the SatMAE model for encoded feature extraction to obtain latent features. These latent features are then input into the predictor, which captures long-term dependencies through multi-layer temporal location encoding and multi-head self-attention, outputting the NDVI prediction sequence for the next three years. ;

[0038] During the prediction period, the mean squared error and slope consistency loss are used for joint training to ensure that the fitted values ​​also maintain the trend direction. The prediction results are then processed by a discrimination threshold that is from the same source as the historical slope to obtain the future trend direction and potential category.

[0039] A second aspect of the present invention also provides a roof greening trend identification and prediction system based on deep time-series modeling. The system includes a memory and a processor. The memory includes a roof greening trend identification and prediction program based on deep time-series modeling. When executed by the processor, the roof greening trend identification and prediction program based on deep time-series modeling performs the following steps:

[0040] S1: Obtain Sentinel-2 NDVI time-series images for a preset number of years, extract building outlines and construction years using OpenStreetMap, and construct the NDVI time series for each building;

[0041] S2: The SatMAE-NDVI autoencoder model is used for self-supervised pre-training. Temporal features are extracted through the mask reconstruction mechanism, and a multi-class classifier is used to determine the overall trend direction of the model output. In each NDVI time series, the reconstruction error sequence and attention weight are output, and the potential mutation point is detected. The NDVI time series is divided into two segments according to the mutation point, and the segments are input into the lightweight regressor to predict the slope value of each segment.

[0042] S3: Based on the relationship between the slope values ​​before and after the mutation point and the overall trend direction, the greening evolution of each NDVI time series is classified into seven typical patterns.

[0043] S4: After completing the classification of greening evolution, a predictor with Transformer as the core is constructed. The NDVI time series after classification is input for prediction, and the NDVI prediction value for the next three years is output. The predictor uses multi-layer temporal location coding and multi-head self-attention to capture long-term dependencies for sequence prediction.

[0044] A third aspect of the present invention also provides a computer-readable storage medium comprising a roof greening trend identification and prediction program based on deep time-series modeling, wherein when the roof greening trend identification and prediction program based on deep time-series modeling is executed by a processor, it implements the steps of the roof greening trend identification and prediction method based on deep time-series modeling as described in any of the preceding claims.

[0045] This invention proposes a method for identifying and predicting rooftop greening trends based on deep temporal modeling. It constructs an integrated analysis process of "trend identification – mutation detection – trend prediction," combining self-supervised learning and the Transformer structure. Addressing the limitations of traditional methods in temporal modeling capabilities and predictive extension, it offers the following significant advantages:

[0046] Automated Trend Identification at the Architectural Scale: This invention, based on architectural-grade NDVI time series, utilizes the SatMAE-NDVI self-supervised model to automatically determine the direction of greening trends. It avoids the reliance on manual thresholds and linear assumptions inherent in traditional statistical methods, making it suitable for dynamic monitoring tasks of diverse roof structures in high-density cities. This significantly improves the accuracy and automation level of trend identification. Furthermore, it significantly enhances the accuracy of identifying architectural-grade roof greening trends, making it applicable to complex urban surface scenarios with multi-temporal and significantly nonlinear changes.

[0047] Integrated Segmented Trend Modeling and Mutation Identification: This invention automatically locates mutation points through model reconstruction error and an attention mechanism, and performs segmented modeling and trend regression on the sequences before and after the mutation to achieve fine-grained characterization of the change stages. Combining the overall trend direction and the relationship between segmented slopes, seven typical change patterns are constructed, exhibiting high sensitivity and adaptability to temporal structures. The mutation location is jointly identified through model reconstruction error and a self-attention mechanism, and a lightweight regression network is used to predict the trend slope before and after the mutation. This demonstrates strong temporal modeling and resistance to abnormal interference, and can stably reflect the staged greening evolution process.

[0048] Extensibility with Future Trend Prediction Function: Compared to methods that only identify historical trends, this invention introduces a Transformer predictor to achieve extended modeling of NDVI trends over the next three years. It possesses strong time-dependent capture capabilities and non-linear trend learning capabilities, providing forward-looking support for urban greening decisions and resource allocation, and expanding the depth and scope of application of rooftop greening time-series analysis in practical scenarios. It supports complete closed-loop modeling from historical assessment to future prediction, providing clear and traceable structured reference results for urban rooftop greening regulation. Attached Figure Description

[0049] Figure 1 A flowchart of a roof greening trend identification and prediction method based on deep temporal modeling according to the present invention is shown;

[0050] Figure 2 The overall flowchart of the time series modeling and prediction of the present invention is shown;

[0051] Figure 3 The diagram shows a block diagram of a roof greening trend identification and prediction system based on deep time series modeling according to the present invention. Detailed Implementation

[0052] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0054] Figure 1 The flowchart of a method for identifying and predicting roof greening trends based on deep time-series modeling according to the present invention is shown.

[0055] like Figure 1 As shown, the first aspect of this invention provides a method for identifying and predicting rooftop greening trends based on deep time-series modeling, comprising:

[0056] S1: Obtain Sentinel-2 NDVI time-series images for a preset number of years, extract building outlines and construction years using OpenStreetMap, and construct the NDVI time series for each building;

[0057] S2: The SatMAE-NDVI autoencoder model is used for self-supervised pre-training. Temporal features are extracted through the mask reconstruction mechanism, and a multi-class classifier is used to determine the overall trend direction of the model output. In each NDVI time series, the reconstruction error sequence and attention weight are output, and the potential mutation point is detected. The NDVI time series is divided into two segments according to the mutation point, and the segments are input into the lightweight regressor to predict the slope value of each segment.

[0058] S3: Based on the relationship between the slope values ​​before and after the mutation point and the overall trend direction, the greening evolution of each NDVI time series is classified into seven typical patterns.

[0059] S4: After completing the classification of greening evolution, a predictor with Transformer as the core is constructed. The NDVI time series after classification is input for prediction, and the NDVI prediction value for the next three years is output. The predictor uses multi-layer temporal location coding and multi-head self-attention to capture long-term dependencies for sequence prediction.

[0060] It should be noted that in the prediction process of S4, by coupling the output trend slope with the historical trend, it is possible to predict the direction of the subsequent trend, which can effectively assist in urban greening management and resource allocation optimization.

[0061] The method of this invention realizes the process of "multi-temporal NDVI acquisition, deep trend recognition, mutation detection, trend classification, and future prediction", and the system supports greening time series modeling and trend judgment at the building scale.

[0062] Figure 2 The overall flowchart of the time series modeling and prediction of the present invention is shown.

[0063] like Figure 2 As shown, it includes four process modules: 1. Raw image acquisition and preprocessing module; 2. Overall and segmented trend verification; 3. Trend change classification; 4. Future trend prediction.

[0064] The specific process for S1 is as follows: Figure 2 The original image acquisition and preprocessing module is shown in the figure.

[0065] It's worth noting that current methods for identifying the evolution of rooftop greening at the building scale mainly fall into two categories: statistical tests and image classification. The former often employs the Mann-Kendall trend test and Theil-Sen slope estimation to analyze the changing trends of NDVI time-series data, exhibiting a certain degree of robustness. However, it relies on linear assumptions, making it difficult to handle non-monotonic, multi-stage changes, and it lacks the ability to identify abrupt changes. The latter identifies rooftop greening areas through NDVI pixel aggregation combined with building outline Zonal Statistics, or using OBIA classification models based on NDVI and NDBI. While these methods can improve recognition accuracy, in high-density cities, the heterogeneity of roof materials, image distortion, and shadow interference make it difficult to accurately extract dynamic changes, limiting their application in large-scale built-up areas.

[0066] At the same time, existing technologies have many problems, such as simplified trend modeling methods that are difficult to cope with complex evolution. Existing roof greening trend identification methods mostly rely on traditional statistical models such as the Mann-Kendall test and Theil-Sen estimation, which assume that the trend is monotonous and changes steadily, making it difficult to capture nonlinear trends, phased reversals and weak but continuous greening evolution, resulting in insufficient expressive power for the real change patterns.

[0067] Existing technologies lack effective mutation detection mechanisms, making it difficult to classify in a structured manner. Most methods fail to systematically detect potential mutation points in NDVI sequences, lack segmentation processing and trend coupling, making it difficult to identify typical change patterns and thus unable to achieve structured classification and precise monitoring of rooftop greening behavior.

[0068] Existing technologies cannot predict future evolution trends and have limited monitoring effectiveness. Most existing methods remain at the stage of passive analysis of historical data and lack forward-looking prediction capabilities based on deep time series learning. This makes it difficult to predict future greening trends and hinders early intervention and resource allocation.

[0069] This invention effectively addresses the aforementioned problems, resolving issues such as inaccurate dynamic identification of rooftop greening at the building scale, untimely detection of abrupt changes, and lack of future change prediction in existing technologies. It proposes a building-scale rooftop greening trend identification and future prediction method based on deep temporal modeling. This method is based on the SatMAE-NDVI autoencoder model proposed by MIT and NASA in 2025. This model, based on the Masked Autoencoder (MAE) architecture, can automatically extract trend features from multi-temporal NDVI sequences. By introducing a self-supervised reconstruction error and attention mechanism, unsupervised detection of greening abrupt changes and segmented trend discrimination are achieved, and the evolution patterns are divided into seven categories based on slope relationships. Finally, a Transformer-based predictor is constructed to predict NDVI trends for the next three years, realizing a complete process from trend identification to future prediction, providing technical support for the dynamic monitoring and forward-looking management of green infrastructure in high-density cities.

[0070] According to an embodiment of the present invention, S1 specifically includes:

[0071] Sentinel-2 L2A multi-temporal remote sensing images from 2018 to 2024 were retrieved using GEE. All remote sensing images were preprocessed with cloud masking and radiometric normalization. The 95th percentile NDVI value of the year was extracted for each pixel to obtain the pixel-level NDVI.

[0072] Using OpenStreetMap building vector outlines and ArcGIS Zonal Statistics, pixel NDVI is aggregated by building unit to form an annual sequence of length 7 in the time dimension. Combined with the year the building was first constructed Eliminate The observed values, For sequence time points;

[0073] Based on the building outline and the year of construction, an annual NDVI time series is generated for each building.

[0074] It should be noted that GEE refers to the Google Earth Engine platform. The specific S1 process is as follows: Figure 2 The original imagery acquisition and preprocessing are shown in the figure. OSM in the figure refers to OpenStreetMap. Extracting the 95th percentile NDVI value of the year ensures the sensitivity of the sequence to the maximum potential of vegetation. The preprocessing process ensures the logical consistency of the sequence. Executing ArcGIS ZonalStatistics involves using the Zonalstatistics tool in ArcGIS software for data statistics.

[0075] According to an embodiment of the present invention, step S2 specifically includes:

[0076] The NDVI time series data was imported into the SatMAE-NDVI autoencoder model for self-supervised pre-training.

[0077] By employing a mask reconstruction method to learn the latent temporal structure of NDVI sequences, a three-classifier is used to determine the overall trend direction of greening. In each NDVI time series, the mean is calculated based on the sliding window of the model, and the reconstruction error and the location of the mutation point are calculated. A lightweight regressor is used to output the slope values ​​of the segments before and after the mutation point.

[0078] It should be noted that S2 corresponds to Figure 2 The overall and segmented trend testing module uses latent vectors to represent the latent temporal structure.

[0079] The SatMAE-NDVI autoencoder model is specifically the SatMAE-NDVI model proposed by MIT (Massachusetts Institute of Technology) and NASA in 2025. This model is based on the mask autoencoder (MAE) architecture and can automatically extract trend features from multi-temporal NDVI sequences.

[0080] According to an embodiment of the present invention, in step S2, the determination of the overall trend direction specifically includes:

[0081] NDVI time series The input is fed into a SatMAE-NDVI autoencoder model, which is pre-trained on multiple NDVI time series using a mask reconstruction task to obtain latent vectors that can simultaneously represent multi-year trends and seasonal rhythms. During the fine-tuning phase, a three-class classifier is constructed. right Perform discrimination and output strictly defined trend labels. The three-category representation is as follows:

[0082] ,

[0083] in, Indicates the overall trend classification. This indicates three classification results.

[0084] It should be noted that the three classifier is trained using cross-entropy loss, and the labels are derived from historical samples that have been manually reviewed, ensuring the certainty and statistical consistency of the judgment results.

[0085] According to an embodiment of the present invention, step S2 further includes:

[0086] For each building sequence, re-perform mask reconstruction of the SatMAE-NDVI autoencoder model and calculate the time-by-time reconstruction error. Calculate the mean within a sliding window of length w. ,when When the value exceeds the mean of the entire sequence plus three standard deviations, that moment is fixed as the unique mutation point. ,by The building sequence is divided into two parts, with the first part serving as the boundary. With the latter part And input them into the lightweight slope regressor respectively. Output and The details are as follows:

[0087] ,

[0088] ,

[0089] in, and All of these are output slope values.

[0090] It should be noted that the slope regressor uses a large number of synthetic sequences with controllable slopes as the training set, and learns the slope mapping through mean squared error loss to ensure that the output value is consistent with the physical rate of change.

[0091] It should be noted that w is the preset window length.

[0092] According to an embodiment of the present invention, S3 specifically includes:

[0093] Based on the sign relationship between the overall trend label and the slopes of the two segments, the greening evolution of the building sequence is classified into seven typical trend types, as follows:

[0094] Continued to rise: , , ;

[0095] Continued decline: , , ;

[0096] It first rises and then falls, but the overall trend is upward: , , ;

[0097] It first rose and then fell, resulting in an overall decline. , , ;

[0098] First decrease, then increase, overall rise: , , ;

[0099] First decrease, then increase, overall decrease: , , ;

[0100] No clear trend: The overall trend is not significant.

[0101] It should be noted that the building sequence is the NDVI time series.

[0102] The classification process employs hard threshold decision-making to avoid introducing subjective weights, ensuring the certainty of the results and guaranteeing that trend judgments are structured and interpretable. The seven typical trend type labels are: Consistent Rise, Consistent Fall, Rise then Fall (Overall Rise), Rise then Fall (Overall Fall), Fall then Rise (Overall Rise), Fall then Rise (Overall Fall), and No Clear Trend.

[0103] The specific process for S3 is as follows: Figure 2 The trend change classification module is shown.

[0104] According to an embodiment of the present invention, S4 specifically includes:

[0105] Build a Transformer-based predictor;

[0106] After obtaining the classification results of the seven typical S3 patterns, the complete NDVI time series is input into the SatMAE model for encoded feature extraction to obtain latent features. These latent features are then input into the predictor, which captures long-term dependencies through multi-layer temporal location encoding and multi-head self-attention, outputting the NDVI prediction sequence for the next three years. ;

[0107] During the prediction period, the mean squared error and slope consistency loss are used for joint training to ensure that the fitted values ​​also maintain the trend direction. The prediction results are then processed by a discrimination threshold that is from the same source as the historical slope to obtain the future trend direction and potential category.

[0108] It should be noted that this invention introduces an NDVI predictor based on the Transformer architecture. Based on latent features extracted from historical sequence encoding, it predicts the NDVI change trajectory over the next three years, assisting in the forward-looking prediction of rooftop greening trends and providing direct quantitative evidence for urban rooftop greening intervention and resource allocation. (SatMAE model)

[0109] The specific process for S4 is as follows: Figure 2 The future trend prediction module is shown below.

[0110] According to an embodiment of the present invention, it further includes:

[0111] Obtain the NDVI prediction sequence for each building;

[0112] Randomly select the prediction sequence of N buildings and mark it as the current prediction sequence;

[0113] Based on a current prediction sequence, analyze the prediction sequences of multiple neighboring buildings within the corresponding preset range to obtain multiple neighbor prediction sequences;

[0114] Based on the grey relational analysis method, the correlation degree between the current predicted sequence and the neighboring predicted sequences is calculated, and multiple correlation degrees are obtained;

[0115] The correlation degree is specifically calculated by taking the current predicted sequence and the neighboring predicted sequences as the reference sequence and the comparison sequence, respectively, using the NDVI value in the sequence as the sequence point for difference statistics, calculating the absolute difference of each sequence point, determining the minimum difference and the maximum difference between the reference sequence and the comparison sequence, and finally calculating the correlation degree based on the grey relational analysis method.

[0116] Obtain multiple distance values ​​between multiple neighboring buildings and the selected building;

[0117] Based on linear regression, calculate the linear fitting coefficients between multiple correlation degrees and multiple building distance values, and set the prediction error value based on the linear fitting coefficients;

[0118] During the predictor training process, error averaging with prediction error value and double loss is introduced for prediction training, and predictor parameters are adjusted in real time.

[0119] It should be noted that the preset years are 2018–2024. The double loss refers to the mean squared error and the slope consistency error. The linear fitting coefficient is inversely proportional to the prediction error value. The larger the linear fitting coefficient, the more correlated the building sequence data and building distance are between the neighboring building sequence and the selected central building sequence in the prediction data, indicating that the prediction conforms to a certain degree of actual changes in building greening, and the smaller the corresponding prediction error should be set. A selected building has multiple neighboring buildings; therefore, there are multiple correlations and multiple [predictions / errors]. When randomly selecting buildings, the average prediction error value can be calculated based on multiple selections of building objects for error introduction training.

[0120] It is worth mentioning that in high-density urban areas with high-density green buildings, theoretical predictions based on data often suffer from overfitting, especially for building clusters with high building density and growing green space. The changes in greening within these building clusters are often correlated. Therefore, this invention introduces correlation analysis and error assessment between the prediction sequence and building distance. This effectively trains and corrects the predictor based on real change patterns, ensuring that the greening prediction sequence is highly consistent with the actual situation, improving temporal stability, and maintaining the accuracy of prediction data even under conditions of dense urban buildings and complex roof shapes. This provides strong data support for monitoring the status of rooftop greening at the building scale.

[0121] Figure 3 The diagram shows a block diagram of a roof greening trend identification and prediction system based on deep time series modeling according to the present invention.

[0122] A second aspect of the present invention also provides a roof greening trend identification and prediction system 3 based on deep time-series modeling. The system includes a memory 31 and a processor 32. The memory includes a roof greening trend identification and prediction program based on deep time-series modeling. When the processor executes the roof greening trend identification and prediction program based on deep time-series modeling, it performs the following steps:

[0123] S1: Obtain Sentinel-2 NDVI time-series images for a preset number of years, extract building outlines and construction years using OpenStreetMap, and construct the NDVI time series for each building;

[0124] S2: The SatMAE-NDVI autoencoder model is used for self-supervised pre-training. Temporal features are extracted through the mask reconstruction mechanism, and a multi-class classifier is used to determine the overall trend direction of the model output. In each NDVI time series, the reconstruction error sequence and attention weight are output, and the potential mutation point is detected. The NDVI time series is divided into two segments according to the mutation point, and the segments are input into the lightweight regressor to predict the slope value of each segment.

[0125] S3: Based on the relationship between the slope values ​​before and after the mutation point and the overall trend direction, the greening evolution of each NDVI time series is classified into seven typical patterns.

[0126] S4: After completing the classification of greening evolution, a predictor with Transformer as the core is constructed. The NDVI time series after classification is input for prediction, and the NDVI prediction value for the next three years is output. The predictor uses multi-layer temporal location coding and multi-head self-attention to capture long-term dependencies for sequence prediction.

[0127] When the system runs S1~S4, it implements all the steps of the above-described method for identifying and predicting roof greening trends based on deep time series modeling.

[0128] A third aspect of the present invention also provides a computer-readable storage medium comprising a roof greening trend identification and prediction program based on deep time-series modeling, wherein when the roof greening trend identification and prediction program based on deep time-series modeling is executed by a processor, it implements the steps of the roof greening trend identification and prediction method based on deep time-series modeling as described in any of the preceding claims.

[0129] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0130] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0131] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0132] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0134] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying and predicting rooftop greening trends based on deep time-series modeling, characterized in that, include: S1: Obtain Sentinel-2 NDVI time-series images for a preset number of years, extract building outlines and construction years using OpenStreetMap, and construct the NDVI time series for each building; S2: The SatMAE-NDVI autoencoder model is used for self-supervised pre-training. Temporal features are extracted through the mask reconstruction mechanism, and a multi-class classifier is used to determine the overall trend direction of the model output. In each NDVI time series, the reconstruction error sequence and attention weight are output, and the potential mutation point is detected. The NDVI time series is divided into two segments according to the mutation point, and the segments are input into the lightweight regressor to predict the slope value of each segment. S3: Based on the relationship between the slope values ​​before and after the mutation point and the overall trend direction, the greening evolution of each NDVI time series is classified into multiple modes. S4: After completing the classification of greening evolution, a predictor with Transformer as the core is constructed. The NDVI time series after classification is input for prediction, and the NDVI prediction value for the next three years is output. The predictor uses multi-layer temporal location coding and multi-head self-attention to capture long-term dependencies for sequence prediction.

2. The method for identifying and predicting rooftop greening trends based on deep time-series modeling as described in claim 1, characterized in that, Specifically, S1 is: Sentinel-2 L2A multi-temporal remote sensing images from 2018 to 2024 were retrieved using GEE. All remote sensing images were preprocessed with cloud masking and radiometric normalization. The 95th percentile NDVI value of the year was extracted for each pixel to obtain the pixel-level NDVI. Using OpenStreetMap building vector outlines, ArcGIS Zonal Statistics was executed to aggregate pixel NDVI by building unit, forming an annual sequence of length 7 in the time dimension. Combined with the year the building was first constructed Eliminate Observed values; Based on the building outline and the year of construction, an annual NDVI time series is generated for each building.

3. The method for identifying and predicting rooftop greening trends based on deep time-series modeling according to claim 2, characterized in that, Specifically, S2 is: The NDVI time series data was imported into the SatMAE-NDVI autoencoder model for self-supervised pre-training. By employing a mask reconstruction method to learn the latent temporal structure of NDVI sequences, a three-classifier is used to determine the overall trend direction of greening. In each NDVI time series, the mean is calculated based on the sliding window of the model, and the reconstruction error and the location of the mutation point are calculated. A lightweight regressor is used to output the slope values ​​of the segments before and after the mutation point.

4. The method for identifying and predicting rooftop greening trends based on deep time-series modeling according to claim 3, characterized in that, In S2, the overall trend direction determination is specifically as follows: NDVI time series The input is fed into a SatMAE-NDVI autoencoder model, which is pre-trained on multiple NDVI time series using a mask reconstruction task to obtain latent vectors that can simultaneously represent multi-year trends and seasonal rhythms. During the fine-tuning phase, a three-class classifier is constructed. right Perform discrimination and output strictly defined trend labels. The three-category representation is as follows: ， in, Indicates the overall trend classification, This indicates three classification results.

5. The method for identifying and predicting rooftop greening trends based on deep time-series modeling according to claim 4, characterized in that, The S2 further includes: For each building sequence, re-perform mask reconstruction of the SatMAE-NDVI autoencoder model and calculate the time-by-time reconstruction error. Calculate the mean within a sliding window of length w. ,when When the value exceeds the mean of the entire sequence plus three standard deviations, that moment is fixed as the unique mutation point. ,by The building sequence is divided into two parts, with the first part serving as the boundary. With the latter part And input them into the lightweight slope regressor respectively. Output and The details are as follows: ， ， in, and All of these are output slope values.

6. The method for identifying and predicting rooftop greening trends based on deep time-series modeling according to claim 5, characterized in that, Specifically, S3 is: Based on the sign relationship between the overall trend label and the slopes of the two segments, the greening evolution of the building sequence is classified into seven typical trend types, as follows: Continued to rise: , , ; Continued decline: , , ; It first rises and then falls, but the overall trend is upward: , , ; It first rose and then fell, resulting in an overall decline. , , ; First decrease, then increase, overall rise: , , ; First decrease, then increase, overall decrease: , , ; No clear trend: The overall trend is not significant.

7. The method for identifying and predicting rooftop greening trends based on deep time-series modeling according to claim 6, characterized in that, Specifically, S4 is: Build a Transformer-based predictor; After obtaining the classification results of the seven typical S3 patterns, the complete NDVI time series is input into the SatMAE model for encoded feature extraction to obtain latent features. These latent features are then input into the predictor, which captures long-term dependencies through multi-layer temporal location encoding and multi-head self-attention, outputting the NDVI prediction sequence for the next three years. ; During the prediction period, the mean squared error and slope consistency loss are used for joint training to ensure that the fitted values ​​also maintain the trend direction. The prediction results are then processed by a discrimination threshold that is from the same source as the historical slope to obtain the future trend direction and potential category.

8. A rooftop greening trend identification and prediction system based on deep time-series modeling, characterized in that, The system includes a memory and a processor. The memory includes a roof greening trend identification and prediction program based on deep time-series modeling. When the processor executes the roof greening trend identification and prediction program based on deep time-series modeling, it performs the following steps: S1: Obtain Sentinel-2 NDVI time-series images for a preset number of years, extract building outlines and construction years using OpenStreetMap, and construct the NDVI time series for each building; S2: The SatMAE-NDVI autoencoder model is used for self-supervised pre-training. Temporal features are extracted through the mask reconstruction mechanism, and a multi-class classifier is used to determine the overall trend direction of the model output. In each NDVI time series, the reconstruction error sequence and attention weight are output, and the potential mutation point is detected. The NDVI time series is divided into two segments according to the mutation point, and the segments are input into the lightweight regressor to predict the slope value of each segment. S3: Based on the relationship between the slope values ​​before and after the mutation point and the overall trend direction, the greening evolution of each NDVI time series is classified into seven typical patterns. S4: After completing the classification of greening evolution, a predictor with Transformer as the core is constructed. The NDVI time series after classification is input for prediction, and the NDVI prediction value for the next three years is output. The predictor uses multi-layer temporal location coding and multi-head self-attention to capture long-term dependencies for sequence prediction.

9. A rooftop greening trend identification and prediction system based on deep time-series modeling as described in claim 8, characterized in that, Specifically, S1 is: Sentinel-2 L2A multi-temporal remote sensing images from 2018 to 2024 were retrieved using GEE. All remote sensing images were preprocessed with cloud masking and radiometric normalization. The 95th percentile NDVI value of the year was extracted for each pixel to obtain the pixel-level NDVI. Using OpenStreetMap building vector outlines, ArcGIS Zonal Statistics was executed to aggregate pixel NDVI by building unit, forming an annual sequence of length 7 in the time dimension. Combined with the year the building was first constructed Eliminate Observed values; Based on the building outline and the year of construction, an annual NDVI time series is generated for each building.

Images